Open a video app and it seems to know what might hold your attention. It does not read your mind. It estimates probabilities from traces of behaviour: what you watched, skipped, searched for, saved or clicked, and how similar items performed for other people. Those clues are imperfect, but at scale they can make the next recommendation feel uncannily personal.¹
First, find a manageable shortlist
A large platform may hold millions of items. It cannot examine every one in depth for every refresh. Recommendation systems often work in stages. One stage retrieves a set of plausible candidates, using signals such as your recent activity, item features and patterns among users with overlapping interests. Another stage ranks that smaller set for the current moment. Google's published description of YouTube recommendations illustrates this candidate-generation-and-ranking design; Meta describes comparable multi-stage machinery for Instagram Explore.¹ ²
The ranker converts features into a score for a chosen goal. That goal might combine predicted clicks, viewing time, likes, survey feedback, freshness or other measures. The exact recipe varies by service and changes over time. “What you might click” is therefore only one part of a wider design decision about what the platform wants to show.
What a click can and cannot say
A click is a convenient training signal because it is easy to count. But people click for many reasons: interest, anger, curiosity, confusion, even an accidental touch. Watching to the end does not necessarily mean enjoying something, while quickly leaving might mean the answer appeared immediately. Platforms increasingly use several signals rather than treating one behaviour as a perfect vote.³
Models can also learn relationships humans would struggle to write as rules. You may never have watched a particular creator, but the topic, pacing, language and viewers' responses may resemble things you have enjoyed. Such patterns are statistical. They can be useful without proving that the system understands you or the content.
The feedback loop
Recommendations change behaviour, and that behaviour becomes new training data. If a system shows you three cooking videos and you click one, it may show you more. That can help you discover a new interest, or it can narrow your feed until it overstates a passing mood. A feed is partly a mirror of past actions and partly a product of what the system decided to offer.²
Designers face a trade-off between familiar choices and exploration. Showing only what previously worked risks repetition. Showing too much unfamiliar material risks irrelevance. New creators and new users pose a further problem because the system has little history on which to rely. Context—time, device, location where appropriate—may help, but it also raises questions about privacy and consent.
Prediction is not judgement
The system can be very good at predicting the next click and still be poor at judging what benefits a person. A sensational headline may attract attention without informing. A useful article may be saved for later and never clicked immediately. The platform's optimisation target shapes which of those outcomes it rewards.³
How a model learns
During training, a system is shown examples of past impressions and responses. It adjusts its internal parameters so that its predicted responses come closer to what happened. Later, it applies that learned pattern to candidates for a particular user. The process can capture subtle combinations—someone who likes short travel videos at lunch may respond differently to a long documentary in the evening—but a learned correlation is not a statement about why the person behaved that way.¹
The system also needs guardrails around the prediction. Items may be filtered for age suitability, duplication or policy reasons. Ranking can include diversity so that a feed does not repeat one creator endlessly. These choices are editorial and product decisions implemented in software, even if a model supplies the scores.
What users can influence
A deliberate follow, search or “not interested” action often gives a clearer clue than passive scrolling, though the platform decides how strongly to weigh it. Clearing a watch history, switching accounts or changing settings may change recommendations, but the effects vary. Some systems react quickly to a new topic; others retain older patterns for longer.
It is easy to overread a single strange recommendation. It may have been a test of unfamiliar material, a popular item shown to many people, or the consequence of a weak statistical association. Recommendation systems are designed to be useful on average, not to provide a transparent account of each choice. Their errors are particularly visible because we experience a feed as if it were tailored to us personally.
That is why recommendations should be understood as guesses within a designed environment, not an objective map of your preferences. The next card in a feed is the result of candidate retrieval, scoring and feedback. You supply signals, but the platform decides which signals matter and which possibilities you get to see.